For normal loads, fully eliminate the load. For the TFE case, adjust
the dmask value in the instruction so the selector doesn't need to
handle it. For the TFE special case, I guess it would be possible to
replace the loaded data register with undef, but as-is this will start
treating it as a well defined value.
Trim elements that won't be written. The equivalent still needs to be
done for writes. Also start widening 3 elements to 4
elements. Selection will get the count from the dmask.
We were letting G_ANYEXT with a vcc register bank through, which was
incorrect and would select to an invalid copy. Fix this up like G_ZEXT
and G_SEXT. Also drop old code to fixup the non-boolean case in
RegBankSelect. We now have to perform that expansion during selection,
so there's no benefit to doing it during RegBankSelect.
Fixes integers that don't evenly divide to i32 pieces. We should
probably extract some of the code in the legalizer to start handling
argument breakdowns. I'm dissatisfied with the argument lowering's
handling of vectors for example, and we should not be producing the
weird G_EXTRACTs we do now.
I used the implementation for floor instead of round. It also turns
out the OpenCL builtin library wasn't using the round builtin, but
implemented the expanded form.
Follow-up for D74433
What the function returns are almost standard BFD names, except that "ELF" is
in uppercase instead of lowercase.
This patch changes "ELF" to "elf" and changes ARM/AArch64 to use their BFD names.
MIPS and PPC64 have endianness differences as well, but this patch does not intend to address them.
Advantages:
* llvm-objdump: the "file format " line matches GNU objdump on ARM/AArch64 objects
* "file format " line can be extracted and fed into llvm-objcopy -O literally.
(https://github.com/ClangBuiltLinux/linux/issues/779 has such a use case)
Affected tools: llvm-readobj, llvm-objdump, llvm-dwarfdump, MCJIT (internal implementation detail, not exposed)
Reviewed By: jhenderson
Differential Revision: https://reviews.llvm.org/D76046
Summary:
Truncating the result of a merge means that most likely we could have done without merge in the first place and just used the input merge inputs directly. This can be done in three cases:
1. If the truncation result is smaller than the merge source, we can use the source in the trunc directly
2. If the sizes are the same, we can replace the register or use a copy
3. If the truncation size is a multiple of the merge source size, we can build a smaller merge
This gets rid of most of the larger, hard-to-legalize merges.
Reviewers: qcolombet, aditya_nandakumar, aemerson, paquette, arsenm, Petar.Avramovic
Reviewed By: arsenm
Subscribers: sdardis, jvesely, wdng, nhaehnle, rovka, jrtc27, atanasyan, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75915
Followup to D75114, this patch reuses the existing MatchRotate ROTL/ROTR rotation pattern code to also recognize the more general FSHL/FSHR funnel shift patterns when we have variable shift amounts, matched with MatchFunnelPosNeg which acts in an (almost) equivalent manner to MatchRotatePosNeg.
Only collapse inner endcf if the outer one belongs to SI_IF.
If it does belong to SI_ELSE then mask being restored in fact
a partial inverse of what we need.
Differential Revision: https://reviews.llvm.org/D76154
There are some functional regressions and I suspect our
scopes are not as perfectly enclosed as I expected.
Disable it for now.
Differential Revision: https://reviews.llvm.org/D76148
Currently we infer whether the flat-scratch-init kernel input should
be enabled based on calls. Move this handling, so we can decide if the
full set of ABI inputs is needed in kernels. Ideally we would have an
analysis of some sort, rather than the function attributes.
The patch removes late endcf handling and only leaves the
related portion with redundant exec mask copy elimination.
Differential Revision: https://reviews.llvm.org/D76095
This patch allows ISD::FSHR(i32) patterns to lower to ALIGNBIT instructions.
This improves test coverage of ISD::FSHR matching - x86 has both FSHL/FSHR instructions and we prefer FSHL by default.
Differential Revision: https://reviews.llvm.org/D76070
This is to replace the optimization from the SIOptimizeExecMaskingPreRA.
We have less opportunities in the control flow lowering because many
VGPR copies are still in place and will be removed later, but we know
for sure an instruction is SI_END_CF and not just an arbitrary S_OR_B64
with EXEC.
The subsequent change needs to convert s_and_saveexec into s_and and
address new TODO lines in tests, then code block guarded by the
-amdgpu-remove-redundant-endcf option in the pre-RA exec mask optimizer
will be removed.
Differential Revision: https://reviews.llvm.org/D76033
Instead, emit a trap and a warning. We force inlining of this
situation, so any function where this happens should be dead as
indirect or external calls are not yet supported. This should avoid
erroring on dead code.
The assumption is that conditional regions are perfectly nested
and a mask restored at the exit from the inner block will be
completely covered by a mask restored in the outer.
It turns out with our current structurizer this is not always
the case.
Disable the optimization for now, but I want to keep it around
for a while to either try after further structurizer changes or
to move it into control flow lowering where we have more info
and reuse the test.
Differential Revision: https://reviews.llvm.org/D75958
Summary:
There's a lot of test case churn but the overall effect is to increase
the number of back-to-back v_sub,v_subbrev pairs, which can execute with
no delay even on gfx10.
Reviewers: arsenm, rampitec, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75999
This was failng on any pre-assigned copy to the VCC bank.
This is something of a workaround for the default implementation in
getInstrMappingImpl, and how it treats copy-like operations in
general.
Copy-like operations are considered to only have one result register
bank, rather than separate banks for each source like a normal
instruction. To avoid potentially mishandling reg_sequence with
impossible operand combinations, the generic implementation errors on
impossible costs. If the bank was already assigned, is treated it
as-if it were an unsatisfiable REG_SEQUENCE mapping. We really don't
get any value from any of what getInstrMappingImpl tries to do for
copies, so just directly emit the simple mapping we really want.
If the loaded memory size was smaller than the result size, this would
produce out of bounds memory accesses. I'm wondering if we need a
distinct narrow memory legalize action type, since a case I care about
is decomposing a 4-byte unaligned access into 4 extending loads, which
would leave the original result register type. I'm currently awkwardly
using narrowScalar to handle unaligned accesses that need to be split.
This avoids regressions in a future patch. I'm confused by the use of
the gfx9 usage legacy_mad. Was this a pointless instruction rename, or
uses fmul_legacy handling? Why is regular mac avilable in that case?
In case the source value ends up in a VGPR, insert a readfirstlane to
avoid producing an illegal copy later. If it turns out to be
unnecessary, it can be folded out.
Summary:
When SI_INDIRECT_DST_V* pseudos has indexes in VGPR, they get expanded into the self-looped basic block that modifies EXEC in a loop.
To keep EXEC consistent it is stored before and then re-stored after the pseudo expansion result.
%95:vreg_512 = SI_INDIRECT_DST_V16 %93:vreg_512(tied-def 0), %94:sreg_32, 0, killed %1500:vgpr_32
results to
s_mov_b64 s[6:7], exec
BB0_16:
v_readfirstlane_b32 s8, v28
v_cmp_eq_u32_e32 vcc, s8, v28
s_and_saveexec_b64 vcc, vcc
s_set_gpr_idx_on s8, gpr_idx(DST)
v_mov_b32_e32 v6, v25
s_set_gpr_idx_off
s_xor_b64 exec, exec, vcc
s_cbranch_execnz BB0_16
; %bb.17:
s_mov_b64 exec, s[6:7]
The bug appeared in case this expansion occurs in the ELSE block of the CF.
Originally
%110:vreg_512 = SI_INDIRECT_DST_V16 %103:vreg_512(tied-def 0), %85:vgpr_32, 0, %107:vgpr_32,
%112:sreg_64 = SI_ELSE %108:sreg_64, %bb.19, 0, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
expanded to
****************** <== here exec has "THEN" context
s_mov_b64 s[6:7], exec
BB0_16:
v_readfirstlane_b32 s8, v28
v_cmp_eq_u32_e32 vcc, s8, v28
s_and_saveexec_b64 vcc, vcc
s_set_gpr_idx_on s8, gpr_idx(DST)
v_mov_b32_e32 v6, v25
s_set_gpr_idx_off
s_xor_b64 exec, exec, vcc
s_cbranch_execnz BB0_16
; %bb.17:
s_or_saveexec_b64 s[4:5], s[4:5] <-- exec mask is restored for "ELSE" but immediately overwritten.
s_mov_b64 exec, s[6:7]
The rest of the "ELSE" block is executed not by the workitems which constitute the "else mask" but by those which constitute "then mask"
SILowerControlFlow::emitElse always considers the basic block begin() as an insertion point for s_or_saveexec.
Proposed fix: The SI_INDIRECT_DST_V* procedure should split the reminder block to create landing pad for the EXEC restoration.
Reviewers: rampitec, vpykhtin, nhaehnle
Reviewed By: vpykhtin
Subscribers: arsenm, kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75472
When expanding scalar packed operations, we should not introduce
illegal vector casts LegalizerHelper introduces. We're not in a
legalizer context, and there's no RegBankSelect apply or legalize
worklist.
We weren't considering the packed case correctly, and this was passing
through to the selector. The selector only checked the size, so this
would incorrectly compile to a single 32-bit scalar add.
As usual, the LegalizerHelper is somewhat awkward to use from
applyMappingImpl. I think this is the first place we've needed
multi-step legalization here though.
Most importantly, this fixes ret i8. Also make sure to handle
signext/zeroext for odd types > i32. Some of the corresponding
argument passing fixes also need to be handled.
Interpret these as extending to the next multiple of 32-bits. This had
no effect with i48 for example, which is really split into {i32, i16},
which should extend the high part.
Use a hack to only enable this for GlobalISel.
Technically this also works with SelectionDAG, but the divergence
selection isn't reliable enough and a few cases fail, but I have no
desire to spend time writing the manual expansion code for it. The DAG
actually does a better job since it catches using v_add_lshl_u32 in
the mixed SGPR/VGPR cases.
This reverts commit 5583c2f2fb.
The lldb bot failure was a test that was fragile and sensitive to irrelevant
changes in instruction ordering. Re-committing this as the test should have
been skipped for AArch64 now.
Differential Revision: https://reviews.llvm.org/D75555
The new behavior matches GNU objdump. A pair of angle brackets makes tests slightly easier.
`.foo:` is not unique and thus cannot be used in a `CHECK-LABEL:` directive.
Without `-LABEL`, the CHECK line can match the `Disassembly of section`
line and causes the next `CHECK-NEXT:` to fail.
```
Disassembly of section .foo:
0000000000001634 .foo:
```
Bdragon: <> has metalinguistic connotation. it just "feels right"
Reviewed By: rupprecht
Differential Revision: https://reviews.llvm.org/D75713
This changes the output of `llvm-readelf -n` from:
```
Displaying notes found at file offset 0x<...> with length 0x<...>:
```
to:
```
Displaying notes found in: .note.foo
```
And similarly, adds a `Name:` field to the `llvm-readobj -n` output for notes.
This change not only increases GNU compatibility, it also makes it much easier to read notes. Note that we still fall back to printing the file offset/length in cases where we don't have a section name, such as when printing notes in program headers or printing notes in a partially stripped file (GNU readelf does the same).
Fixes llvm.org/PR41339.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D75647
Summary:
This seems like an obvious error - cut and paste issue?
The change does make a change to one of the lit tests - it stops s_buffer_load
re-ordering past an MUBUF instruction (which is not surprising).
Change-Id: I80be99de5b62af4f42e91af2591b76a52ac9efa6
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75686
After structurization, some phi nodes can have a single incoming edge
and can be simplified away. This change runs a simplify query on all
phis that are either modified or added by the structurizer. This also
moves some phis closer to their use as a side benefit.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D75500
Create a wider source vector, and unmerge with dead defs like the
legalizer. The legalization handling for G_EXTRACT is incomplete, and
it's preferrable to keep everything in 32-bit pieces.
We should probably start moving these functions into utils, since we
have a growing number of places that do almost the same thing.
Since this is still largely relying on the DAG argument type lowering
code, this has inherited the problem where i16 vectors have a
different ABI on targets with and without legal i16. Switch to using a
target with legal i16, so the i16 vector argument tests are more
useful.
If SimplifyDemandedBits succeeds in simplifying the byte src, add the CVT_F32_UBYTE node back to the worklist as we might be able to simplify further.
Yet another step towards removing SelectionDAG::GetDemandedBits.
This changes the localizer to attempt intra-block localizer of instructions
that have local uses. This is useful because sometimes the entry block itself
has many uses of constant-like instructions, which would benefit from shortening
live ranges. Previously if an inst had no non-local uses, we wouldn't add it to
the list of instructions to attempt further intra-block localization.
This gives a 0.7% geomean code size improvement on CTMark.
Differential Revision: https://reviews.llvm.org/D75555
The computation here didn't really make sense to me, and reported
wildy different results depending on the flat work group size
attribute.
I think this should really report a range derived from the possible
work group size bounds, and only allow an occupancy that is a multiple
of the group size.
This is needed for D74873, AMDGPU going to have 16 bit subregs
and the largest tuple is 32 VGPRs, which results in 64 lanes.
Differential Revision: https://reviews.llvm.org/D75378
Summary:
pickNodeBidirectional tried to compare the best top candidate and the
best bottom candidate by examining TopCand.Reason and BotCand.Reason.
This is unsound because, after calling pickNodeFromQueue, Cand.Reason
does not reflect the most important reason why Cand was chosen. Rather
it reflects the most recent reason why it beat some other potential
candidate, which could have been for some low priority tie breaker
reason.
I have seen this cause problems where TopCand is a good candidate, but
because TopCand.Reason is ORDER (which is very low priority) it is
repeatedly ignored in favour of a mediocre BotCand. This is not how
bidirectional scheduling is supposed to work.
To fix this I changed the code to always compare TopCand and BotCand
directly, like the generic implementation of pickNodeBidirectional does.
This removes some uncommented AMDGPU-specific logic; if this logic turns
out to be important then perhaps it could be moved into an override of
tryCandidate instead.
Graphics shader benchmarking on gfx10 shows a lot more positive than
negative effects from this change.
Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB
Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68338
Summary:
The point of this is to make some tests with manual checks robust
against scheduler tweaks, so that only autogenerated test updates will
be required when pushing D68338 "[AMDGPU] Remove dubious logic in
bidirectional list scheduler".
Reviewers: arsenm, rampitec, vpykhtin
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75302
Summary:
The affected LIT test intends to test the correct use of divergence
analysis to detect a divergent branch with a uniform predicate. The
passes involved are LLVM IR passes, but the test runs llc and tries to
match against generated ISA, which makes it hard to demonstrate that
the intended behavior was really tested. Replaced this with a test
that invokes opt on the required passes and then checks for the
appropriate changes in the LLVM IR.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D75267
Summary:
The old code made some incorrect assumptions about the order in which
basic blocks are laid out in a function. This could lead to incorrect
early-exits, especially when kills occurred inside of loops.
The new approach is to check whether the point where the conditional
kill occurs dominates all reachable code. If that is the case, there
cannot be any other threads in the wave that are waiting to rejoin
at a later point in the CFG, i.e. if exec=0 at that point, then all
threads really are dead and we can exit the wave.
Make some other minor cleanups to the pass while we're at it.
v2: preserve the dominator tree
Reviewers: arsenm, cdevadas, foad, critson
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74908
Change-Id: Ia0d2b113ac944ad642d1c622b6da1b20aa1aabcc
The current set of custom combines are only really useful after
legalization, so move them there. There is a lot of overlap in the
boilerplate here, but I think we do want a pretty different set of
combines before and after legalize. I think we will want a lot of
overlap between the post-legalize and a post-regbankselect combiner.
Changes the handling of odd breakdowns, and avoids using
G_EXTRACT/G_INSERT. Pad with undef to a wider size, and unmerge. Also
avoid introducing instructions for the fully undef components.
This reverts commit 977cd661cf.
It breaks OpenCL testing. OpenCL Runtime is using PT_LOAD information
to calculate memory for global variables. This commit should be relanded once
the OpenCL runtime stops relying on PT_LOAD information for calculating global
variable memory size.
Differential Revision: https://reviews.llvm.org/D74995
The legalizer helper functions are unusably awkward to perform the 3-5
part legalization. This needs to be widened, scalarized, lowered, and
we should avoid creating vector extends and truncates. Manually do all
of this and expand.
I tried to use some of the new tablegen features to avoid creating
different operand list permutations, but I still don't see a way to
programmatically build a source pattern dag.
Also add GlobalISel tests, which now all import successfully.
Some of the fneg fold tests are incorrect, which need to be fixed in a
future commit
G_SHUFFLE_VECTOR is legal since it theoretically may help match op_sel
for VOP3P instructions. Expand it in some other way in case it doesn't
fold into the use instructions.
We should try the generated matchers before the manual selection. This
means the patterns are now handling the common cases, but the manual
selection code is not yet dead. It's still handling the non-s32/s64
cases (like v2s16 and v2s32). Currently tablegen doesn't have a nice
way to have a single pattern that covers multiple types.
We have patterns for s_pack* selection, but they assume the inputs are
a build_vector with 16-bit inputs, not a truncating build
vector. Since there's still outstanding work for how to handle
mismatched result and source element vector operations, and since I'm
trying a different packed vector strategy than SelectionDAG, just
manually select this for now.
There are few differences from the DAG handling. First, the DAG
handling uses a primitive selection pattern instead of custom
legalizing it. Because of this, this makes use of source modifiers
while the DAG does not.
Also instead of promoting f16, try to use the f16 log/exp. There's no
f16 fmul_legacy, so widen just for the multiply, although I'm not sure
that's the best solution.
This looked through copies to find the source modifiers, which may
have been SGPR->VGPR copies added to avoid potential constant bus
violations. Re-insert a copy to a VGPR if this happens.
Similar to what we already do with SimplifyDemandedVectorElts, call SimplifyDemandedBits across all the extracted elements of the source vector, treating it as single use.
There's a minor regression in store-weird-sizes.ll which will be addressed in an upcoming SimplifyDemandedBits patch.
Marking a section as ALLOC tells the ELF loader to load the section into memory.
As we do not want to load the notes into VRAM, the flag should not be there.
Differential Revision: https://reviews.llvm.org/D74600
GetDemandedBits mostly just calls SimplifyMultipleUseDemandedBits now, but it does a very blunt constant simplification that SimplifyMultipleUseDemandedBits avoids.
If we need to demand bits from constants we should handle this through ShrinkDemandedConstant/targetShrinkDemandedConstant.
@arsenm confirmed that the sign extended immediates are better for code size.
Differential Revision: https://reviews.llvm.org/D74857
Summary:
After updating cost model in AMDGPU target (47a5c36b37) the pass started to
ignore some BBs since they got all instructions estimated as free.
Reviewers: arsenm, chandlerc, nhaehnle
Reviewed By: nhaehnle
Subscribers: jvesely, wdng, nhaehnle, tpr, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74825
We probably want this, and I've meant to turn this on for a long
time. SC actually emits a special case to early-out for a 1
denominator, which perhaps should also be considered.
Use common check prefix, although update_mir_test_checks makes this
unnecessarily annoying. Also make sure to have uses in case that ever
ends up mattering.
This is part of the work to remove SelectionDAG::GetDemandedBits and just use SimplifyMultipleUseDemandedBits.
Recent experiments raised some v_cvt_f32_ubyte*_e32 regressions, so I've added some additional abilities to performCvtF32UByteNCombine to help unpack byte data more aggressively.
We still don't remove all OR(SHL,SRL) patterns as some of the regenerated nodes don't get combined again, but we are getting closer.
Differential Revision: https://reviews.llvm.org/D74786
AMDGPUCodeGenPrepare expands this most of the time, but not always. We
will always at least need a fallback option here. This is the 3rd
implementation of the same expansion in the backend. Eventually I
would like to eliminate the IR expansion (and the DAG version
obviously).
Currently the new legalizer path produces a better result, since the
IR expansion results in extra operations which need to be combined
out. Notably, the IR expansion results in multiplies by 0.
Produce an unmerge to a narrower type and introduce a narrower shift
if needed. I wasn't sure if there was a better way to parameterize the
target's preferred shift type for the GICombineRule, so manually call
the combine helper.
While looking at the output on real sized programs, there is a lot of
extra SGPR spilling compared to the DAG path. This seems to largely be
from all constants being SGPRs in the entry block.
Try to handle arbitrary scalar BFEs by packing the operands. The DAG
gives up on non-constant arguments. We're still missing any constant
folding, so we end up with pretty ugly code most of the time. Also
handle the 64-bit scalar case, which the DAG doesn't try to do.
Summary:
This fixes some tests that did not specify -mcpu. Doing that disables
all subtarget features, which gives behavior that (a) does not
necessarily correspond to any actual target, and (b) can change as we
add new subtarget features.
Also added gfx1010 to memtime test.
Differential Revision: https://reviews.llvm.org/D74594
Change-Id: I8c0fe4fa03e9a93ef8bb722cd42d22e064526309
The division expansions in AMDGPUCodeGenPrepare can't be relied on for
correctness, since they punt to later optimization and possibly
legalization in some cases. We still need a way to be able to write
tests for the legalizer versions of the expansion. This is mostly for
GlobalISel, since the expected optimzations is expecting aren't
implemented.
The interaction with the flag to expand 64-bit division in the IR is
pretty confusing, but these flags have different purposes.
I didn't realize we were already expanding 24/32-bit division here
already. Use the available IntegerDivision utilities. This uses loops,
so produces significantly smaller code than the inline DAG expansion.
This now requires width reductions of 64-bit divisions before
introducing the expanded loops.
This helps work around missing legalization in GlobalISel for
division, which are the only remaining core instructions that didn't
work at all.
I think this is plausibly a better implementation than exists in the
DAG, although turning it on by default misses out on the constant
value optimizations and also needs benchmarking.
This is more or less directly ported from the AMDGPU custom lowering
for FP_TO_FP16. I made a few minor fixups (using G_UNMERGE_VALUES
instead of creating shift/trunc to extract the two halves, and zexting
an inverted compare instead of select_cc).
This also does not include the fast math expansion the DAG which
converts to f32 and then to f16. I think that belongs in a
pre-legalize combine instead.
This reverts commit 80a34ae311 with fixes.
Previously, since bots turning on EXPENSIVE_CHECKS are essentially turning on
MachineVerifierPass by default on X86 and the fact that
inline-asm-avx-v-constraint-32bit.ll and inline-asm-avx512vl-v-constraint-32bit.ll
are not expected to generate functioning machine code, this would go
down to `report_fatal_error` in MachineVerifierPass. Here passing
`-verify-machineinstrs=0` to make the intent explicit.
This reverts commit 80a34ae311 with fixes.
On bots llvm-clang-x86_64-expensive-checks-ubuntu and
llvm-clang-x86_64-expensive-checks-debian only,
llc returns 0 for these two tests unexpectedly. I tweaked the RUN line a little
bit in the hope that LIT is the culprit since this change is not in the
codepath these tests are testing.
llvm\test\CodeGen\X86\inline-asm-avx-v-constraint-32bit.ll
llvm\test\CodeGen\X86\inline-asm-avx512vl-v-constraint-32bit.ll
Also greatly improve i64 lowering. LegalizeIntegerTypes does the
correct narrowing if i64 isn't legal. Just workaround this for
SelectionDAG by making i64 legal and splitting in the patterns.
When SI_IF is inserted, it constrains the source register with a
register class, which was quite likely a G_ICMP. This was incorrectly
treating it as a scalar, and then applyMappingImpl would end up
producing invalid MIR since this was unexpected.
Also fix not using all VGPR sources for vcc outputs.
These are only legal on CI+. The test would fail in a debug build, but
not a release due to the partial selection since the pre-selection
legality assert only happens in a debug build.
Load extra bits if suitably aligned. This allows using widened
3-vector loads on SI, and fixes legalization for <9 x s32> (which LSV
apparently forms frequently on lowered kernel argument lists).
Fix incorrectly treating these as legal on SI. This should emit a
64-bit store and a 32-bit store.
I think all of the load and store rules are just about complete, but
due for a rewrite.
Summary:
The method attempts to find loads that can be legally clustered by
looking for loads consuming the same chain glue token.
However, the old code looks at _all_ users of values produced by the
chain node -- including uses of the loaded/returned value of volatile
loads or atomics. This could lead to circular dependencies which then
failed during scheduling.
With this change, we filter out users by getResNo, i.e. by which
SDValue value they use, to ensure that we only look at users of the
chain glue token.
This appears to be a rather old bug, which is perhaps surprising.
However, the test case is actually quite fragile (i.e., it is hidden
by fairly small changes), and the test _must_ use volatile loads for
the bug to manifest.
Reviewers: arsenm, bogner, craig.topper, foad
Subscribers: MatzeB, jvesely, wdng, hiraditya, javed.absar, jfb, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74253
This reverts commit rGcd5b308b828e, rGcd5b308b828e, rG8cedf0e2994c.
There are issues to be investigated for polly bots and bots turning on
EXPENSIVE_CHECKS.
This was creating a select on true/false values, and then comparing
that later. This produced more work for later combines, which can be
avoided by just using the boolean values. This was copied from the
original DAG expansion, which also has the same problem. This doesn't
have a observable change using SelectionDAG, but since GlobalISel is
missing these optimizations, the final code was noticeably longer.
These have nicer expansions implemented in the DAG. Ideally we would
either directly implement all of these special expansions, or stop
expanding division in the IR.
This is apparently worse than 1-byte alignment. This does not attempt
to decompose 2-byte aligned wide stores, but will stop trying to
produce them.
Also fix bug in LoadStoreVectorizer which was decreasing the alignment
and vectorizing stack accesses. It was assuming a stack object was an
alloca that could have its base alignment changed, which is not true
if the pointer is derived from a function argument.
Since natural fdiv lowering is now more conservative even with
denormals disabled, we get a slower expansion from just a plain
1.0/fdiv. Directly emit the rcp intrinsic when using it to implement
integer division to avoid a pointlessly complex sequence.
Add a simplification to fuse a manual vector extract with shifts and
truncate into a bitcast.
Unpacking and packing values into vectors is only optimized with
extractelement instructions, not when manually unpacked using shifts
and truncates.
This patch simplifies shifts and truncates into a bitcast if possible.
Simplify (build_vec (trunc $1)
(trunc (srl $1 width))
(trunc (srl $1 (2 * width))) ...)
to (bitcast $1)
Differential Revision: https://reviews.llvm.org/D73892
Based on D72931
This adds a new feature called A16 which is enabled for gfx10.
gfx9 keeps the R128A16 feature so it can share all the instruction encodings
with gfx7/8.
Differential Revision: https://reviews.llvm.org/D73956
Vector indexing with a constant index should be folded out in the
legalizer, but this was accidentally falling through. This would
produce the indexing operation with $noreg. Handle this case as a
dynamic index just in case a bug like this happens again in the
future.
We were failing to find constants that were casted. I feel like the
artifact combiner should have folded the constant in the trunc before
the custom lowering, but that doesn't happen.
Reverts part of 6524a7a2b9. Since that
commit, the expansion was ignoring the actual save exec register
produced by the instruction, and looking at other instructions. I do
not understand why it was looking at other instructions, but relying
on this scan was wrong.
Fixes verifier errors after SI_IF is tail duplicated, which should be
correct to do. The results were fed into a phi, which was lowered to
the S_MOV_B64_term instructions.
Summary:
For CTTZ we place a set bit just past where the non-promoted type
stopped so the extended bits won't be used for the count. For
CTTZ_ZERO_UNDEF we don't care what happens if no bits are set in
the original type and we end up counting into the extended bits.
So we can just use ANY_EXTEND for both cases.
This matches what is done in type legalization for these operations.
We make no effort to force the upper bits to zero.
Differential Revision: https://reviews.llvm.org/D74111
Summary:
The accuracy limit to use rcp is adjusted to 1.0 ulp from 2.5 ulp.
Also, afn instead of arcp is used to allow inaccurate rcp to be used.
Reviewers:
arsenm
Differential Revision: https://reviews.llvm.org/D73588
Summary:
Current implementation of matchSwap in SIShrinkInstructions searches the entire
use_nodbg_operands set to find the possible pattern to generate v_swap instruction.
This approach will lead to a O(N^3) in compile time for SIShrinkInstructions.
But in reality, the matching pattern only exists within nearby instructions in the
same basic block. This work limits the search to a maximum of 16 instructions, and has
a linear compile time comsumption.
Reviewers:
rampitec, arsenm
Differential Revision: https://reviews.llvm.org/D74180
We were executing this in a waterfall loop as a placeholder, but this
should really be converted to a MUBUF load. Also execute in a
waterfall loop if the resource isn't an SGPR. This is a case where the
DAG handling was wrong because doing the right thing was too hard.
Currently, this will mishandle 96-bit loads. There's currently no way
to track the original memory size with an MMO, so these loads will be
widened andd the resulting memory size will be 128-bits.
The type passed to lower was invalid, so I'm not sure how this was
even working before. The source and destination type also do not have
to match, so make sure to use the right ones.
Summary:
Current implementation of matchSwap in SIShrinkInstructions searches the entire
use_nodbg_operands set to find the possible pattern to generate v_swap instruction.
This approach will lead to a O(N^3) in compile time for SIShrinkInstructions.
But in reality, the matching pattern only exists within nearby instructions in the
same basic block. This work limits the search to a maximum of 16 instructions, and has
a linear compile time comsumption.
Reviewers:
rampitec, arsenm
Differential Revision: https://reviews.llvm.org/D74180
The legalizer produces a lot of these, and they make reading legalized
MIR annoying. For some reason, this does seem to sometimes introduce
copies of implicit def, which is dumb.
Really the intrinsic definition is wrong, but work around this
here. The DAG lowering introduces an MMO. We have to introduce a new
operation to avoid the verifier complaining about the missing mayLoad.
Use cmp ord instead of cmp_class compared to the DAG version for the
nan check, but mostly try to match the existsing pattern.
I think the sign doesn't matter for fract, so we could do a little
better with the source modifier matching.
I think this is also still broken as in D22898, but I'm leaving it
as-is for now while I don't have an SI system to test on.
Rewrite the result register pair into the expected sinigle register
format in the legalizer.
I'm also operating under the assumption that TFE doesn't apply to
stores or atomics, but don't know if this is true or not.
The 96-bit results need to be widened.
I find the interaction between LegalizerHelper and MIRBuilder somewhat
awkward. The custom legalization is called by the LegalizerHelper, but
then does not have access to the helper. You have to construct a new
helper, which then does not own the MachineIRBuilder, but does modify
it. Maybe custom legalization should be passed the helper?
If we have s_pack_* instructions, legalize this to
G_BUILD_VECTOR_TRUNC from s32 elements. This is closer to how how the
s_pack_* instructions really behave.
If we don't have s_pack_ instructions, expand this by creating a merge
to s32 and bitcasting. This expands to the expected bit operations. I
think this eventually should go in a new bitcast legalize action type
in LegalizerHelper.
We already directly emit the shift operations in RegBankSelect for the
vector case. This could possibly be cleaned up, but I also may want to
defer doing this expansion to selection anyway. I'll see about that
when I try to actually match VOP3P instructions.
This breaks the selection of the build_vector since tablegen doesn't
know how to match G_BUILD_VECTOR_TRUNC yet, so just xfail it for now.
scalar_to_vector takes only one argument, not two.
The a16 tests now also check the packing of coordinates into registers
Differential Revision: https://reviews.llvm.org/D73482
This should lower the amount of used registers for gfx9.
I updated some of the changed tests with the update script because
changing them by hand is tedious.
Differential Revision: https://reviews.llvm.org/D73884
The old version might be faster on EG (RECIP_IEEE is Trans only),
but it'd need extra corner case checks.
This gives correct corner case behaviour and saves a register.
Fixes OCL CTS sqrt test (1-thread, scalar) on Turks.
Reviewer: arsenm
Differential Revision: https://reviews.llvm.org/D74017
https://reviews.llvm.org/D72312 introduced an infinite loop which involves
DAGCombiner::visitFMA and AMDGPUTargetLowering::performFNegCombine.
fma( a, fneg(b), fneg(c) ) => fneg( fma (a, b, c) ) => fma( a, fneg(b), fneg(c) ) ...
This only breaks with types where 'isFNegFree' returns flase, e.g. v4f32.
Reproducing the issue also needs the attribute 'no-signed-zeros-fp-math',
and no source mods allowed on one of the users of the Op.
This fix makes changes to indicate that it is not free to negate a fma if it
has users with source mods.
Differential Revision: https://reviews.llvm.org/D73939
Prepare to accurately track the future denormal-fp-math attribute
changes. The way to actually set these separately is not wired in yet.
This is just a mechanical change, and mostly still assumes the input
and output mode match. This should be refined for some cases. For
example, fcanonicalize lowering should use the flushing variant if
either input or output flushing is enabled
Sebastian Neubauer